Re: (weak) benchmarks

[Brad Spencer <brad%anduin.eldar.org@localhost>]

Greg Troxel <gdt%lexort.com@localhost> writes:

> I often hear that xen is slow, pv is slow, and various other things, so
> I'm trying to measure instead of entering lore in to the howto.  If you
> have real data, especially if it's better or better organized than this,
> or if you want to suggest the benchmark from pkgsrc that I should have
> run instead, please do!
>
> My test is simply "make package clean" in devel/m4 (from 2024Q2),
> without ccache, and with WRKDIR on tmpfs /tmp.  The dom0 has disk images
> in zfs.  This is run 4 times but you can see they are all close.  Memory
> sizes vary, but I do not believe any of these are running out of RAM
> during these tests.
>
> This is a pre-2010 "Q6600" processor, that does have VMX.  A 2010
> "E5700" without VMX appeared similar.
>
> The dom0 is normal pv.
> The i386 domUs are pvh/pvshim.
> The amd64 are pv.
>
> It remains to try for domUs
>   pvh
>   hvm
> and for dom0
>   pvh
> to see how that is, both functionally and speed.   pvh/hvm domUs did not
> work with the E5700.
>
> ** baremetal amd64 10 8GB 2024Q2
>
>        81.82 real        55.51 user        24.08 sys
>        81.46 real        56.09 user        23.67 sys
>        81.43 real        55.28 user        24.47 sys
>        81.47 real        55.41 user        24.37 sys
>
> ** dom0 4GB 2024Q2
>
>       161.38 real        71.82 user        84.81 sys
>       160.98 real        72.16 user        84.75 sys
>       161.11 real        71.21 user        85.85 sys
>       161.15 real        72.44 user        84.63 sys
>
> ** domU i386 9 1GB 2024Q2
>
>       225.29 real       114.57 user       102.74 sys
>       231.95 real       119.89 user       105.01 sys
>       233.48 real       121.01 user       105.18 sys
>       230.41 real       118.90 user       104.48 sys
>
> ** domU amd64 9 3GB 2024Q2
>
>       195.93 real        82.20 user       106.28 sys
>       195.34 real        81.62 user       107.12 sys
>       195.67 real        80.88 user       108.17 sys
>       195.36 real        82.74 user       106.34 sys
>
> ** domU i386 10 1GB 2024Q2
>
>       221.40 real       116.13 user        98.72 sys
>       226.58 real       120.22 user       100.78 sys
>       225.94 real       118.69 user       101.83 sys
>       226.57 real       119.32 user       101.71 sys
>
> ** domU amd64 10 3GB 2024Q2
>
>       207.45 real        86.78 user       118.86 sys
>       206.47 real        86.58 user       119.16 sys
>       206.80 real        86.81 user       119.18 sys
>       206.67 real        86.94 user       118.73 sys


This is probably a very weak addition:

The build system is a amd64 pure PVH DOMU running NetBSD 10.0 with 12GB
of memory and 2 vcpus.  The source for 2024Q1 pkgsrc is on SSD as is the
WRKOBJDIR (i.e. the object files and etc...).  However,
/usr/pkgsrc/packages is on spinning rust disks.

28.9u 14.5s 0:57.43 75.7% 4615+1826k 39+31763io 2885pf+0w


This is just "time make package clean" from a tcsh.



52.44 real        29.15 user        14.35 sys

That is "/usr/bin/time make package clean"


The DOM0 is a PV DOM0 running NetBSD 9.3_STABLE with HZ=1000.




-- 
Brad Spencer - brad%anduin.eldar.org@localhost - KC8VKS - http://anduin.eldar.org